We describe the investigation to find more reliable way to recognize speakers in the field. As primary features we use instant formants frequencies (for frames, where formants exist) and average for utterance pitch. In comparing two utterances modified nearest neighbour distance is used. It is discovered that 1ms inter-frame shift gives some noticeable advantage in recognition score. For verification task (30 speakers*6times during 4months) this algorithm showed 2-6% of errors in noisy environment.
Bibliographic reference. Labulin, Pavel V. / Koval, Sergey L. / Raev, Andrej N. (1995): "Automatic speaker recognition using formants-based nearest-neighbour distance measure", In EUROSPEECH-1995, 341-344.